Openpilot 0.10
17 days ago
- #autonomous-driving
- #machine-learning
- #openpilot
- openpilot 0.10 introduces a new training architecture, maintaining reliability while transitioning towards openpilot 1.0.
- New Driving Model features include a new end-to-end training architecture (Tomb Raider), improved Lead Car detection (Space Lab), and VAE compression for training (Vegan Filet-o-Fish).
- World Model Planner replaces MPC systems, improving longitudinal control in Experimental mode and enabling future enhancements.
- Space Lab improves stopped lead car detection, reducing ignored low-speed frames from 78% to 52%.
- VAE compression validates negligible loss in image quality, essential for efficient generative model training.
- Live lateral delay estimation is now enabled, improving steering accuracy on some cars.
- Community driving feedback is streamlined with a new Discord bot for easier bug reporting and video clip generation.
- New features include a clipping tool for video clips and opt-in dashcam audio recording.
- commaCarSegments v2 is released, offering 3000 hours of CAN data from 20k openpilot users.
- Car interface slimming reduces the workload for porting and maintaining car interfaces, with CAN parser rewritten in Python.
- Bug fixes and enhancements include improvements for Hyundai CAN FD, Ford Bronco Sport, and support for new car models like Acura MDX 2025 and Honda Accord 2023-25.
- Community engagement is encouraged through Discord, driving feedback, and development contributions.